A realist’s guide to deploying AI agents that work

AI agents are everywhere. Boards want them. Founders pitch them. Vendors promise them. VCs love them. And they are real and buildable today.

But the promised agent, the one that learns your business, works unsupervised, and makes decisions, is still a few chapters away – despite what any consultant may say.

What we call AI agents are flowcharts glued to large language models (LLMs). They can do useful things, but they don’t understand context, strategy, or judgment.

The smart move for any company looking to use AI agents is to level up responsibly. This means deploying what works, building the right habits, and getting your teams ready for what’s coming next.

As OpenAI co-founder Andrej Karpathy put it: “Don’t build Iron Man robots. Build Iron Man suits.” Simply, build tools that make your people stronger instead of tech that tries to replace them.

Robots and humans have limitations

Most of the AI agents available today are overhyped. While they’re fine for low-order work, they break when tasks get complex.

In my experience with agents today, I see three fundamental issues.

First, by default, they’re brittle, meaning a slight change in input can make the entire chain of action fall apart. Whether it’s a different format, a longer prompt, or a new variable, these changes expose how agents today struggle with multistep reasoning and are unable to adapt.

For example, if a firm is using an agent for asset management work, and a vendor provides data for returns as basis points instead of percentages like normal, chaos could ensue. The agent might run the math without noticing or flagging the change and produce a polished report with completely incorrect results.

See also: Wanted: First movers for Southeast Asia’s agentic AI leap

Also, LLMs still hallucinate, meaning they confidently make things up. If you’re just brainstorming with ChatGPT, that might be OK, but it’s more problematic when its for high-level decision-making.

Finally, agents lack a strategy memory, meaning they don’t learn across tasks and can’t “remember” how your business works. Even Karpathy says we need to “keep AI on a leash” for now.

Aside from these tech issues, the truth is that most companies lack the maturity to take advantage of the benefits that agents do bring. I see three flaws most often.

A blueprint for agent success

Decade, not year